CODEE — AI Coding Assistant

CODEE is a full-stack AI coding assistant built with:

• Qwen 2.5 (0.5B & 3B) fine-tuned models
• Node.js/Express backend with safety guardrails
• React frontend with a modern, responsive UI

It provides safe, high-quality code generation, real-time feedback, automatic refinement via a critic model, and a clean developer experience.

---

🌟 Features

🔧 Backend (Node.js + Qwen)

• Multi-layer prompt validation guardrails
• Code generation with Qwen 2.5 Coder
• “Critic” system for improved, cleaner code
• Structured REST API (/generate-code, /validate)
• Logging, error handling, environment configuration

💻 Frontend (React)

• Dark-themed UI
• Real-time status steps: Validation → Generation → Criticism
• Side-by-side generated and improved code
• One-click copy blocks
• Mobile-responsive layout

📚 Research Notebooks

• SFT for Qwen 2.5 0.5B
• SFT for Qwen 2.5 3B
• Tools for dataset preparation, evaluation, and experimentation

---

📁 Monorepo Structure

CODEE/
├── backend/             # Express API server
│   ├── src/
│   ├── .env.example
│   └── package.json
│
├── frontend/            # React web UI
│   ├── src/
│   ├── .env.example
│   └── package.json
│
└── notebooks/           # SFT, evaluation, dataset prep
    ├── CODEE_qwen_2_5_0_5_SFT.ipynb
    ├── CODEE_qwen_2_5_3_SFT.ipynb
    └── README.md

---

⚙️ Installation & Setup

You can run both services locally for development.

---

1️⃣ Backend Setup (Node + Express)

cd backend
npm install
cp .env.example .env

Edit .env:

HUGGING_FACE_TOKEN=your_token_here
PORT=5000
NODE_ENV=development

Run server:

npm run dev   # dev mode (nodemon)
# or
npm start     # production

Backend runs at:

http://localhost:5000

---

2️⃣ Frontend Setup (React)

cd frontend
npm install
cp .env.example .env

Set backend API URL:

REACT_APP_API_BASE_URL=http://localhost:5000/api
REACT_APP_DEBUG=false

Run client:

npm start

Frontend runs at:

http://localhost:3000

---

👨‍🔬 Notebooks (Model R&D)

Located in /notebooks.

Included:

✔ CODEE_qwen_2_5_0_5_SFT.ipynb

Lightweight 0.5B model fine-tuning notebook → Perfect for prototyping, fast iterations

✔ CODEE_qwen_2_5_3_SFT.ipynb

Large 3B model fine-tuning notebook → Same structure as 0.5B SFT, but with more parameters and better reasoning

These notebooks include:

• Dataset preparation
• Prompt/response formatting
• Supervised fine-tuning training loops
• Loss/evaluation plots
• Inference and model export tools

Run notebooks with:

jupyter lab
# or
jupyter notebook

---

🌐 API Endpoints

Endpoint	Method	Description
/api/generate-code	POST	Validate → Generate → Critic improve
/api/validate	POST	Run guardrails without generating
/health	GET	Backend status

Example request:

{ "prompt": "Write a Python function to reverse a string" }

---

🚢 Deployment

Static Frontend (Netlify, Vercel, GitHub Pages, S3)

npm run build

Backend (Render, Railway, Docker)

Example Docker run:

docker build -f docker/Dockerfile.backend -t coding-assistant-backend .
docker run -p 5000:5000 -e HUGGING_FACE_TOKEN=your_token coding-assistant-backend

---

🧪 Testing

This project includes comprehensive test suites for both frontend and backend.

Backend Tests

cd codee-backend
npm test              # Run all tests
npm run test:coverage # Run with coverage
npm run test:watch    # Watch mode

Note: Backend tests require a .env.test file. Copy from .env.example and set HUGGING_FACE_TOKEN.

Frontend Tests

cd codee-frontend
npm test              # Interactive mode
npm test -- --coverage --watchAll=false  # CI mode with coverage

CI/CD

All tests run automatically on every push and pull request via GitHub Actions:

• Backend CI: Jest tests with ES modules, Prettier formatting checks, coverage reports
• Frontend CI: React Testing Library tests, ESLint checks, build verification, coverage reports
• Unified CI: Runs both in parallel with status checks

Required GitHub Secret:

• HUGGING_FACE_TOKEN - Add this to your repository secrets for backend tests to pass

---

🐞 Troubleshooting

❌ Frontend can’t reach backend

• Backend not running
• Wrong REACT_APP_API_BASE_URL
• CORS issues

⏱ Slow first request

HF models cold-start; repeated calls are faster.

⚠️ Guardrail rejections

Triggered for non-coding and unsafe prompts.

---

📄 License

MIT